Harder, better, faster, stronger: updating our protein folding base miner

Miners have set a blistering pace on SN25, upgrading to GPUs in order to optimize their performance. We're making updates to the subnet to match.

When we launched SN25 in June 2024, our underlying goal was to prove Bitttensor’s capacity for confronting computationally demanding scientific challenges like protein folding simulations. And the subnet’s miners have risen to the challenge. Already, over 400,000 protein folding jobs have been completed on SN25. 

With the core concept proven, we can focus on driving improvements. That’s why we’ve now updated our base machine to support CUDA-supported GPUs as standard.

Miner-craft

When designing Subnet 25, we aimed to place as few limitations as possible on how miners were contributing to the subnet. Protein folding as a discipline utilizes a standardized software package called GROMACS, which miners have to work with. In order to run GROMACS and contribute to the subnet, they required a Linux-based machine and multiple high-performance CPU cores.

As well as setting those parameters, we provide the base miner, which sets a floor for performance on the subnet. That base miner is usually the starting point for new miners looking to start on the subnet.

CPU vs GPU

Since launch, we’ve seen that top-performing miners have assembled increasingly complex GPU architectures to better simulate the conditions that surround protein folding. That’s not surprising; there’s a growing body of literature benchmarking CPU and GPU performance running GROMACS. Protein folding is computationally demanding, requiring miners to repeatedly run calculations that simulate the laws of physics until they have reached an optimal configuration for a folded protein. Facing these requirements, it’s natural that top-performing miners prefer GPU cores.

Top performing miners set the benchmark as to what is possible in Bittensor. As they continue to push the upper limits and compete with each other, innovation is inevitable. With their transition to GPUs and subsequent performance improvements, it was time that we updated the base miner to this paradigm. 

The future: balancing access and performance

For miners, there are still substantial opportunities to improve above the base model. Currently, there are bottlenecks for miners when they’re running multiple processes on their GPUs, slowing down their simulations. Process handling, rather than underlying hardware, has become the key parameter where miners can significantly improve their performance. 

For the subnet as a whole, we also want to maintain that balance between having a low-enough barrier-to-entry that SN25 is accessible to a broad pool of miners, while also ensuring the subnet as a whole is delivering results. That means faster protein folding simulations for end users, and a more competitive ecosystem for miners: a win-win scenario, and a boost for the Bittensor ecosystem, too.